Oral: Interpretable multi-task inference of virus microbe interaction networks from population dynamicsPoster: Interpretable multi-task inference of virus microbe interaction networks from population dynamics

A central goal in microbial and viral ecology is understanding the interactions between multiple coexisting virus strains and microbes. Studying which viruses infect which microbial hosts, which microbial groups compete vs. cooperate, and the strength of these interactions is important for perceiving the intertwined ecology of these coexisting species. However, the lack of mechanistic interpretable principles behind their dynamics often limits our understanding of interactions and traits that lead to large-scale species coexistence.

Here, we integrated multiple tasks for inverse modeling – generalized Lotka-Volterra models and prior parameters from ecology, species stability from nonlinear dynamics, and network regularization from machine learning – to infer these ecological connections from population time series data. This made our inverse model more interpretable as it includes multiple objectives in the inference task ("multi-task" inference). Through in-silico experiments, we found that simply fitting a mechanistic model to coexisting virus microbe population dynamics may result in network and traits suggestive of selective strain elimination, over-fitted dynamics, and biologically implausible traits. In contrast, we found that only by including prior knowledge through the multi-task inference can we infer functional traits and interactions that lead to stable coexistence. Moreover, multi-task inference resulted in both better ecological network estimation and lower population dynamics forecasting errors. Our in-silico results emphasize the necessity of incorporating ecological constraints to accurately infer traits, network interactions, and population dynamics, with broader implications for ecosystem modeling and forecasting.